Using Episodic Memory to Aid Machine Learning Performance in Perceptually Aliased Environments
摘要
In this paper, we begin with an agent using an artificial neural network to improve its behavior in an environment with extreme perceptual aliasing. We then improve the performance of using an episodic memory. In particular, a memory of past events allows the agent to distinguish it’s present state from other states that are perceived identically. We describe steps we took to leverage this memory and remove the manual tuning required for the agent to perform in each specific environment.